Moving Object Tracking Based On Superpixel

As one important part of intelligent information processing system,target tracking technology could be sensed in many areas,such as urban intelligent traffic system,Intelligent Vehicles,and detecting anomaly behavior,etc.Therefore it has widely application values.With the rapid development of information technology,in particular in recent years,the target tracking technology advances rapidly,while the precision,veracity and robustness of target tracking have been dramatically improved.However,because of the complexity of real-world situations,target tracking technology still faces various challenges,as objects tracking results will lose caused by the illumination variation,target appearance changes,the interference of background noises,and so on.In this paper,on the basis of summarizing the predecessors' research,super pixel division technology is trying to apply to target tracking technology,to enhance the robustness of target tracking algorithm.The main contents of this paper are as follows:(1)The five most common segmentation algorithms were introduced in this paper,and one valuation methodology based on color histogram was presented.Using this valuation methodology,the features of color histogram were computed firstly.Then,the energy values of pixels color histogram was calculated by the energy function,while the performance of super pixels algorithm was evaluated by energy values.(2)The four super pixels algorithms were evaluated by marginal recall rate,error rate of under segmentation,and segmentation efficiency of super pixels,as well as valuation methodology presented in this paper.The marginal recall rate,error rate of under segmentation,segmentation efficiency and color features of SEEDS super pixels algorithm were superior to other three algorithms,which found through experiments.Therefore,SEEDS was used to carry out the follow-up tracking experiment.(3)The color features of image was researched,and one feature of super pixels based on improving color coherence vector was put forward.It was described pixels by the normalized multi-threshold color coherence vector with abundant color and structure features.(4)Appearance Models were built and updated so as to,tracking the target successfully.In reaction to the failure of tracking algorithm,the confidence of image was gotten from matching the Appearance Models with candidate images by constructed and updated the Appearance Models,and finally the tracking is realized.The experiment showed that this methodology could obtain good performances in lame tracking in the condition of target distorted,target's rotation,scaling and shelter and etc.